Mapping River Salinization Using Airborne Electromagnetic Data and Unsupervised Machine Learning
Abstract
Salinization of freshwater in rivers is a global and growing threat. It affects the river ecosystems, causes a reduction in biodiversity, and eventually compromises river goods and services provided to humans. Climate change and increasing water consumption can worsen the problem. It is, therefore, critical to map and monitor the salinization of rivers. Conventional methods such as hydrological drilling and sampling are time-consuming, expensive, and inefficient. Airborne electromagnetics (EM) has proven useful due to its ability to quickly cover large areas. To interpret the EM data, 1D, 2D, or 3D inversions are typically performed and followed by either a qualitative or quantitative analysis of the inverted conductivity features. This approach has worked well in previous studies. However, performing inversions is a non-trivial task. In recent years, machine learning has gained immense popularity in geoscience and has been successfully applied to many problems. Our study focuses on evaluating the applicability of machine learning to map salinized areas in rivers based on airborne EM data. Our study area is in a groundwater management trial site near Bookpurnong in South Australia. Airborne frequency and time domain data were collected over the Murray River and the adjacent floodplain. We aim to identify salinization by directly clustering the airborne EM data. For the frequency domain EM data, we first normalized the all-frequency measurements at each location by their maximum value to minimize the effect of the bird heights. We then implemented K-means clustering assuming 2, 3, 4 and 5 clusters and successfully identified the salinized areas that are consistent with previous inversion-based studies. For the time domain EM data, we found that low-moment measurements of SkyTEM system contain more information about the river salinization than the high-moment data. We normalized each time domain decay curve by its maximum value and calculated its logarithms. We then applied K-means clustering to the transformed decay curves, and successfully mapped out the salinized areas. Our work demonstrates the efficacy of unsupervised machine learning in quickly mapping river salinization using open source machine learning frameworks (e.g., Scikit-Learn) and free cloud computing environments (e.g., Google Colab).
- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2021
- Bibcode:
- 2021AGUFMIN42B..06S